Cerebras Methods’ monster debut on Thursday did not simply place it amongst tech’s biggest-ever IPOs — it was a crystal clear sign of unstoppable demand for chips to energy AI, as tech giants scramble to search out options to the pricey, sold-out graphics processing models made by Nvidia.
Cerebras closed its first day buying and selling on Wall Avenue with a market cap just under $100 billion, placing it close to the few firms to shut above that mark, resembling Fb-parent Meta and Alibaba. The inventory traded decrease on Friday, its first full day of buying and selling.
Here is what it’s worthwhile to learn about this sizzling Nvidia competitor.
Cerebras makes a special sort of chip than the traditional Nvidia GPU, and it is the dimensions of a dinner plate.
“We construct the largest chips within the semiconductor business,” Cerebras CEO and Co-Founder Andrew Feldman informed CNBC on Squawk Field Thursday. “Massive chips course of extra info in much less time and ship outcomes extra rapidly.”
Till now, Nvidia has been successful the AI chip race as a result of its GPUs function general-purpose workhorses, excelling on the parallel math vital for coaching massive fashions. However we have now arrived on the period of agentic AI, the place inference is vital. Whereas coaching teaches the AI mannequin to be taught from patterns in massive quantities of information, inference makes use of the AI to make choices primarily based on new info.
Inference can occur on much less highly effective chips programmed for extra particular duties, resembling Cerebras’ WSE-3. It falls right into a class of chips generally known as customized ASICs — application-specific built-in circuits. It is an more and more crowded house, with in-house ASICS now made by the likes of Google, Amazon, Meta and Microsoft.
Cerebras stated the WSE-3 is 57 occasions bigger than the most important GPU, and has 50 occasions the variety of transistors.
Probably the most superior AI chips are made utilizing Taiwan Semiconductor Manufacturing’s 2-nanometer course of node, presently solely attainable in Taiwan. Cerebras’ chip can also be made at TSMC, however on its much less superior 5-nanometer node.
Based in Silicon Valley in 2016, Cerebras first filed to go public in 2024, however withdrew that submission when it confronted scrutiny for its heavy reliance on a single buyer, Microsoft-backed AI agency G42 within the United Arab Emirates.
With the agency’s profitable IPO on Thursday, Feldman and {hardware} expertise chief Sean Lie, two of its co-founders, grew to become billionaires primarily based on their holdings.
Cerebras one-week inventory chart.
For years, Cerebras sought to promote chips to firms, however now largely operates the chips inside its personal knowledge facilities as a cloud service, pitting it in opposition to cloud suppliers Google, Microsoft, Oracle and CoreWeave.
Cerebras and OpenAI introduced a $20 billion cloud deal in January that expires in 2028, whereas Amazon Net Companies introduced in March that it is utilizing Cerebras chips in its knowledge facilities.
“For our quick inference product, there’s a lot demand that our largest problem is definitely making an attempt to produce it. We’re including as a lot manufacturing and knowledge heart capability as we probably can, and we’re nonetheless offered out into 2027,” Cerebras CFO Bob Komin informed CNBC Thursday.
Whereas hyperscalers make their very own in-house ASICs, Cerebras extra carefully competes with corporations focusing on making them for others. Chief amongst these is Groq. In its largest buy thus far, Nvidia paid $20 billion for Groq’s tech in December, then introduced customized Groq Language Processing Models at GTC in March.
SambaNova and D-Matrix are two different notable Cerebras opponents trying to capitalize on unprecedented AI chip demand.
SambaNova counts Hugging Face and Meta among the many clients of its SN50 chips, whereas Intel participated in a $350 million funding spherical for SambaNova in February. Intel CEO Lip-Bu Tan has served as SambaNova’s chairman since 2017.
Cerebras’ IPO additionally paves the way in which for different customized ASIC start-ups trying to go public, resembling Rebellions.
The South Korean chipmaker raised $400 million from the likes of Samsung, at a valuation of $2.34 billion, in March because it prepares for an IPO.
Watch: Breaking down AI chips, from Nvidia GPUs to ASICs by Google and Amazon
